1 //===- MemorySSA.cpp - Memory SSA Builder ---------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the MemorySSA class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Analysis/MemorySSA.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/DenseMapInfo.h"
17 #include "llvm/ADT/DenseSet.h"
18 #include "llvm/ADT/DepthFirstIterator.h"
19 #include "llvm/ADT/Hashing.h"
20 #include "llvm/ADT/None.h"
21 #include "llvm/ADT/Optional.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/ADT/SmallPtrSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/iterator.h"
26 #include "llvm/ADT/iterator_range.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/IteratedDominanceFrontier.h"
29 #include "llvm/Analysis/MemoryLocation.h"
30 #include "llvm/Config/llvm-config.h"
31 #include "llvm/IR/AssemblyAnnotationWriter.h"
32 #include "llvm/IR/BasicBlock.h"
33 #include "llvm/IR/CallSite.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/Instruction.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/IntrinsicInst.h"
39 #include "llvm/IR/Intrinsics.h"
40 #include "llvm/IR/LLVMContext.h"
41 #include "llvm/IR/PassManager.h"
42 #include "llvm/IR/Use.h"
43 #include "llvm/Pass.h"
44 #include "llvm/Support/AtomicOrdering.h"
45 #include "llvm/Support/Casting.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Compiler.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/FormattedStream.h"
51 #include "llvm/Support/raw_ostream.h"
60 #define DEBUG_TYPE "memoryssa"
62 INITIALIZE_PASS_BEGIN(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,
64 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
65 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
66 INITIALIZE_PASS_END(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,
69 INITIALIZE_PASS_BEGIN(MemorySSAPrinterLegacyPass, "print-memoryssa",
70 "Memory SSA Printer", false, false)
71 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
72 INITIALIZE_PASS_END(MemorySSAPrinterLegacyPass, "print-memoryssa",
73 "Memory SSA Printer", false, false)
75 static cl::opt<unsigned> MaxCheckLimit(
76 "memssa-check-limit", cl::Hidden, cl::init(100),
77 cl::desc("The maximum number of stores/phis MemorySSA"
78 "will consider trying to walk past (default = 100)"));
81 VerifyMemorySSA("verify-memoryssa", cl::init(false), cl::Hidden,
82 cl::desc("Verify MemorySSA in legacy printer pass."));
86 /// An assembly annotator class to print Memory SSA information in
88 class MemorySSAAnnotatedWriter : public AssemblyAnnotationWriter {
89 friend class MemorySSA;
91 const MemorySSA *MSSA;
94 MemorySSAAnnotatedWriter(const MemorySSA *M) : MSSA(M) {}
96 void emitBasicBlockStartAnnot(const BasicBlock *BB,
97 formatted_raw_ostream &OS) override {
98 if (MemoryAccess *MA = MSSA->getMemoryAccess(BB))
99 OS << "; " << *MA << "\n";
102 void emitInstructionAnnot(const Instruction *I,
103 formatted_raw_ostream &OS) override {
104 if (MemoryAccess *MA = MSSA->getMemoryAccess(I))
105 OS << "; " << *MA << "\n";
109 } // end namespace llvm
113 /// Our current alias analysis API differentiates heavily between calls and
114 /// non-calls, and functions called on one usually assert on the other.
115 /// This class encapsulates the distinction to simplify other code that wants
116 /// "Memory affecting instructions and related data" to use as a key.
117 /// For example, this class is used as a densemap key in the use optimizer.
118 class MemoryLocOrCall {
122 MemoryLocOrCall() = default;
123 MemoryLocOrCall(MemoryUseOrDef *MUD)
124 : MemoryLocOrCall(MUD->getMemoryInst()) {}
125 MemoryLocOrCall(const MemoryUseOrDef *MUD)
126 : MemoryLocOrCall(MUD->getMemoryInst()) {}
128 MemoryLocOrCall(Instruction *Inst) {
129 if (ImmutableCallSite(Inst)) {
131 CS = ImmutableCallSite(Inst);
134 // There is no such thing as a memorylocation for a fence inst, and it is
135 // unique in that regard.
136 if (!isa<FenceInst>(Inst))
137 Loc = MemoryLocation::get(Inst);
141 explicit MemoryLocOrCall(const MemoryLocation &Loc) : Loc(Loc) {}
143 ImmutableCallSite getCS() const {
148 MemoryLocation getLoc() const {
153 bool operator==(const MemoryLocOrCall &Other) const {
154 if (IsCall != Other.IsCall)
158 return Loc == Other.Loc;
160 if (CS.getCalledValue() != Other.CS.getCalledValue())
163 return CS.arg_size() == Other.CS.arg_size() &&
164 std::equal(CS.arg_begin(), CS.arg_end(), Other.CS.arg_begin());
169 ImmutableCallSite CS;
174 } // end anonymous namespace
178 template <> struct DenseMapInfo<MemoryLocOrCall> {
179 static inline MemoryLocOrCall getEmptyKey() {
180 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getEmptyKey());
183 static inline MemoryLocOrCall getTombstoneKey() {
184 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getTombstoneKey());
187 static unsigned getHashValue(const MemoryLocOrCall &MLOC) {
191 DenseMapInfo<MemoryLocation>::getHashValue(MLOC.getLoc()));
194 hash_combine(MLOC.IsCall, DenseMapInfo<const Value *>::getHashValue(
195 MLOC.getCS().getCalledValue()));
197 for (const Value *Arg : MLOC.getCS().args())
198 hash = hash_combine(hash, DenseMapInfo<const Value *>::getHashValue(Arg));
202 static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS) {
207 } // end namespace llvm
209 /// This does one-way checks to see if Use could theoretically be hoisted above
210 /// MayClobber. This will not check the other way around.
212 /// This assumes that, for the purposes of MemorySSA, Use comes directly after
213 /// MayClobber, with no potentially clobbering operations in between them.
214 /// (Where potentially clobbering ops are memory barriers, aliased stores, etc.)
215 static bool areLoadsReorderable(const LoadInst *Use,
216 const LoadInst *MayClobber) {
217 bool VolatileUse = Use->isVolatile();
218 bool VolatileClobber = MayClobber->isVolatile();
219 // Volatile operations may never be reordered with other volatile operations.
220 if (VolatileUse && VolatileClobber)
222 // Otherwise, volatile doesn't matter here. From the language reference:
223 // 'optimizers may change the order of volatile operations relative to
224 // non-volatile operations.'"
226 // If a load is seq_cst, it cannot be moved above other loads. If its ordering
227 // is weaker, it can be moved above other loads. We just need to be sure that
228 // MayClobber isn't an acquire load, because loads can't be moved above
231 // Note that this explicitly *does* allow the free reordering of monotonic (or
232 // weaker) loads of the same address.
233 bool SeqCstUse = Use->getOrdering() == AtomicOrdering::SequentiallyConsistent;
234 bool MayClobberIsAcquire = isAtLeastOrStrongerThan(MayClobber->getOrdering(),
235 AtomicOrdering::Acquire);
236 return !(SeqCstUse || MayClobberIsAcquire);
241 struct ClobberAlias {
243 Optional<AliasResult> AR;
246 } // end anonymous namespace
248 // Return a pair of {IsClobber (bool), AR (AliasResult)}. It relies on AR being
249 // ignored if IsClobber = false.
250 static ClobberAlias instructionClobbersQuery(MemoryDef *MD,
251 const MemoryLocation &UseLoc,
252 const Instruction *UseInst,
254 Instruction *DefInst = MD->getMemoryInst();
255 assert(DefInst && "Defining instruction not actually an instruction");
256 ImmutableCallSite UseCS(UseInst);
257 Optional<AliasResult> AR;
259 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(DefInst)) {
260 // These intrinsics will show up as affecting memory, but they are just
262 switch (II->getIntrinsicID()) {
263 case Intrinsic::lifetime_start:
265 return {false, NoAlias};
266 AR = AA.alias(MemoryLocation(II->getArgOperand(1)), UseLoc);
267 return {AR == MustAlias, AR};
268 case Intrinsic::lifetime_end:
269 case Intrinsic::invariant_start:
270 case Intrinsic::invariant_end:
271 case Intrinsic::assume:
272 return {false, NoAlias};
279 ModRefInfo I = AA.getModRefInfo(DefInst, UseCS);
280 AR = isMustSet(I) ? MustAlias : MayAlias;
281 return {isModOrRefSet(I), AR};
284 if (auto *DefLoad = dyn_cast<LoadInst>(DefInst))
285 if (auto *UseLoad = dyn_cast<LoadInst>(UseInst))
286 return {!areLoadsReorderable(UseLoad, DefLoad), MayAlias};
288 ModRefInfo I = AA.getModRefInfo(DefInst, UseLoc);
289 AR = isMustSet(I) ? MustAlias : MayAlias;
290 return {isModSet(I), AR};
293 static ClobberAlias instructionClobbersQuery(MemoryDef *MD,
294 const MemoryUseOrDef *MU,
295 const MemoryLocOrCall &UseMLOC,
297 // FIXME: This is a temporary hack to allow a single instructionClobbersQuery
298 // to exist while MemoryLocOrCall is pushed through places.
300 return instructionClobbersQuery(MD, MemoryLocation(), MU->getMemoryInst(),
302 return instructionClobbersQuery(MD, UseMLOC.getLoc(), MU->getMemoryInst(),
306 // Return true when MD may alias MU, return false otherwise.
307 bool MemorySSAUtil::defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU,
309 return instructionClobbersQuery(MD, MU, MemoryLocOrCall(MU), AA).IsClobber;
314 struct UpwardsMemoryQuery {
315 // True if our original query started off as a call
317 // The pointer location we started the query with. This will be empty if
319 MemoryLocation StartingLoc;
320 // This is the instruction we were querying about.
321 const Instruction *Inst = nullptr;
322 // The MemoryAccess we actually got called with, used to test local domination
323 const MemoryAccess *OriginalAccess = nullptr;
324 Optional<AliasResult> AR = MayAlias;
326 UpwardsMemoryQuery() = default;
328 UpwardsMemoryQuery(const Instruction *Inst, const MemoryAccess *Access)
329 : IsCall(ImmutableCallSite(Inst)), Inst(Inst), OriginalAccess(Access) {
331 StartingLoc = MemoryLocation::get(Inst);
335 } // end anonymous namespace
337 static bool lifetimeEndsAt(MemoryDef *MD, const MemoryLocation &Loc,
339 Instruction *Inst = MD->getMemoryInst();
340 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
341 switch (II->getIntrinsicID()) {
342 case Intrinsic::lifetime_end:
343 return AA.isMustAlias(MemoryLocation(II->getArgOperand(1)), Loc);
351 static bool isUseTriviallyOptimizableToLiveOnEntry(AliasAnalysis &AA,
352 const Instruction *I) {
353 // If the memory can't be changed, then loads of the memory can't be
356 // FIXME: We should handle invariant groups, as well. It's a bit harder,
357 // because we need to pay close attention to invariant group barriers.
358 return isa<LoadInst>(I) && (I->getMetadata(LLVMContext::MD_invariant_load) ||
359 AA.pointsToConstantMemory(cast<LoadInst>(I)->
360 getPointerOperand()));
363 /// Verifies that `Start` is clobbered by `ClobberAt`, and that nothing
364 /// inbetween `Start` and `ClobberAt` can clobbers `Start`.
366 /// This is meant to be as simple and self-contained as possible. Because it
367 /// uses no cache, etc., it can be relatively expensive.
369 /// \param Start The MemoryAccess that we want to walk from.
370 /// \param ClobberAt A clobber for Start.
371 /// \param StartLoc The MemoryLocation for Start.
372 /// \param MSSA The MemorySSA isntance that Start and ClobberAt belong to.
373 /// \param Query The UpwardsMemoryQuery we used for our search.
374 /// \param AA The AliasAnalysis we used for our search.
375 static void LLVM_ATTRIBUTE_UNUSED
376 checkClobberSanity(MemoryAccess *Start, MemoryAccess *ClobberAt,
377 const MemoryLocation &StartLoc, const MemorySSA &MSSA,
378 const UpwardsMemoryQuery &Query, AliasAnalysis &AA) {
379 assert(MSSA.dominates(ClobberAt, Start) && "Clobber doesn't dominate start?");
381 if (MSSA.isLiveOnEntryDef(Start)) {
382 assert(MSSA.isLiveOnEntryDef(ClobberAt) &&
383 "liveOnEntry must clobber itself");
387 bool FoundClobber = false;
388 DenseSet<MemoryAccessPair> VisitedPhis;
389 SmallVector<MemoryAccessPair, 8> Worklist;
390 Worklist.emplace_back(Start, StartLoc);
391 // Walk all paths from Start to ClobberAt, while looking for clobbers. If one
392 // is found, complain.
393 while (!Worklist.empty()) {
394 MemoryAccessPair MAP = Worklist.pop_back_val();
395 // All we care about is that nothing from Start to ClobberAt clobbers Start.
396 // We learn nothing from revisiting nodes.
397 if (!VisitedPhis.insert(MAP).second)
400 for (MemoryAccess *MA : def_chain(MAP.first)) {
401 if (MA == ClobberAt) {
402 if (auto *MD = dyn_cast<MemoryDef>(MA)) {
403 // instructionClobbersQuery isn't essentially free, so don't use `|=`,
404 // since it won't let us short-circuit.
406 // Also, note that this can't be hoisted out of the `Worklist` loop,
407 // since MD may only act as a clobber for 1 of N MemoryLocations.
408 FoundClobber = FoundClobber || MSSA.isLiveOnEntryDef(MD);
411 instructionClobbersQuery(MD, MAP.second, Query.Inst, AA);
421 // We should never hit liveOnEntry, unless it's the clobber.
422 assert(!MSSA.isLiveOnEntryDef(MA) && "Hit liveOnEntry before clobber?");
424 if (auto *MD = dyn_cast<MemoryDef>(MA)) {
426 assert(!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA)
428 "Found clobber before reaching ClobberAt!");
432 assert(isa<MemoryPhi>(MA));
433 Worklist.append(upward_defs_begin({MA, MAP.second}), upward_defs_end());
437 // If ClobberAt is a MemoryPhi, we can assume something above it acted as a
438 // clobber. Otherwise, `ClobberAt` should've acted as a clobber at some point.
439 assert((isa<MemoryPhi>(ClobberAt) || FoundClobber) &&
440 "ClobberAt never acted as a clobber");
445 /// Our algorithm for walking (and trying to optimize) clobbers, all wrapped up
447 class ClobberWalker {
448 /// Save a few bytes by using unsigned instead of size_t.
449 using ListIndex = unsigned;
451 /// Represents a span of contiguous MemoryDefs, potentially ending in a
455 // Note that, because we always walk in reverse, Last will always dominate
456 // First. Also note that First and Last are inclusive.
459 Optional<ListIndex> Previous;
461 DefPath(const MemoryLocation &Loc, MemoryAccess *First, MemoryAccess *Last,
462 Optional<ListIndex> Previous)
463 : Loc(Loc), First(First), Last(Last), Previous(Previous) {}
465 DefPath(const MemoryLocation &Loc, MemoryAccess *Init,
466 Optional<ListIndex> Previous)
467 : DefPath(Loc, Init, Init, Previous) {}
470 const MemorySSA &MSSA;
473 UpwardsMemoryQuery *Query;
475 // Phi optimization bookkeeping
476 SmallVector<DefPath, 32> Paths;
477 DenseSet<ConstMemoryAccessPair> VisitedPhis;
479 /// Find the nearest def or phi that `From` can legally be optimized to.
480 const MemoryAccess *getWalkTarget(const MemoryPhi *From) const {
481 assert(From->getNumOperands() && "Phi with no operands?");
483 BasicBlock *BB = From->getBlock();
484 MemoryAccess *Result = MSSA.getLiveOnEntryDef();
485 DomTreeNode *Node = DT.getNode(BB);
486 while ((Node = Node->getIDom())) {
487 auto *Defs = MSSA.getBlockDefs(Node->getBlock());
489 return &*Defs->rbegin();
494 /// Result of calling walkToPhiOrClobber.
495 struct UpwardsWalkResult {
496 /// The "Result" of the walk. Either a clobber, the last thing we walked, or
497 /// both. Include alias info when clobber found.
498 MemoryAccess *Result;
500 Optional<AliasResult> AR;
503 /// Walk to the next Phi or Clobber in the def chain starting at Desc.Last.
504 /// This will update Desc.Last as it walks. It will (optionally) also stop at
507 /// This does not test for whether StopAt is a clobber
509 walkToPhiOrClobber(DefPath &Desc,
510 const MemoryAccess *StopAt = nullptr) const {
511 assert(!isa<MemoryUse>(Desc.Last) && "Uses don't exist in my world");
513 for (MemoryAccess *Current : def_chain(Desc.Last)) {
515 if (Current == StopAt)
516 return {Current, false, MayAlias};
518 if (auto *MD = dyn_cast<MemoryDef>(Current)) {
519 if (MSSA.isLiveOnEntryDef(MD))
520 return {MD, true, MustAlias};
522 instructionClobbersQuery(MD, Desc.Loc, Query->Inst, AA);
524 return {MD, true, CA.AR};
528 assert(isa<MemoryPhi>(Desc.Last) &&
529 "Ended at a non-clobber that's not a phi?");
530 return {Desc.Last, false, MayAlias};
533 void addSearches(MemoryPhi *Phi, SmallVectorImpl<ListIndex> &PausedSearches,
534 ListIndex PriorNode) {
535 auto UpwardDefs = make_range(upward_defs_begin({Phi, Paths[PriorNode].Loc}),
537 for (const MemoryAccessPair &P : UpwardDefs) {
538 PausedSearches.push_back(Paths.size());
539 Paths.emplace_back(P.second, P.first, PriorNode);
543 /// Represents a search that terminated after finding a clobber. This clobber
544 /// may or may not be present in the path of defs from LastNode..SearchStart,
545 /// since it may have been retrieved from cache.
546 struct TerminatedPath {
547 MemoryAccess *Clobber;
551 /// Get an access that keeps us from optimizing to the given phi.
553 /// PausedSearches is an array of indices into the Paths array. Its incoming
554 /// value is the indices of searches that stopped at the last phi optimization
555 /// target. It's left in an unspecified state.
557 /// If this returns None, NewPaused is a vector of searches that terminated
558 /// at StopWhere. Otherwise, NewPaused is left in an unspecified state.
559 Optional<TerminatedPath>
560 getBlockingAccess(const MemoryAccess *StopWhere,
561 SmallVectorImpl<ListIndex> &PausedSearches,
562 SmallVectorImpl<ListIndex> &NewPaused,
563 SmallVectorImpl<TerminatedPath> &Terminated) {
564 assert(!PausedSearches.empty() && "No searches to continue?");
566 // BFS vs DFS really doesn't make a difference here, so just do a DFS with
567 // PausedSearches as our stack.
568 while (!PausedSearches.empty()) {
569 ListIndex PathIndex = PausedSearches.pop_back_val();
570 DefPath &Node = Paths[PathIndex];
572 // If we've already visited this path with this MemoryLocation, we don't
573 // need to do so again.
575 // NOTE: That we just drop these paths on the ground makes caching
576 // behavior sporadic. e.g. given a diamond:
581 // ...If we walk D, B, A, C, we'll only cache the result of phi
582 // optimization for A, B, and D; C will be skipped because it dies here.
583 // This arguably isn't the worst thing ever, since:
584 // - We generally query things in a top-down order, so if we got below D
585 // without needing cache entries for {C, MemLoc}, then chances are
586 // that those cache entries would end up ultimately unused.
587 // - We still cache things for A, so C only needs to walk up a bit.
588 // If this behavior becomes problematic, we can fix without a ton of extra
590 if (!VisitedPhis.insert({Node.Last, Node.Loc}).second)
593 UpwardsWalkResult Res = walkToPhiOrClobber(Node, /*StopAt=*/StopWhere);
594 if (Res.IsKnownClobber) {
595 assert(Res.Result != StopWhere);
596 // If this wasn't a cache hit, we hit a clobber when walking. That's a
598 TerminatedPath Term{Res.Result, PathIndex};
599 if (!MSSA.dominates(Res.Result, StopWhere))
602 // Otherwise, it's a valid thing to potentially optimize to.
603 Terminated.push_back(Term);
607 if (Res.Result == StopWhere) {
608 // We've hit our target. Save this path off for if we want to continue
610 NewPaused.push_back(PathIndex);
614 assert(!MSSA.isLiveOnEntryDef(Res.Result) && "liveOnEntry is a clobber");
615 addSearches(cast<MemoryPhi>(Res.Result), PausedSearches, PathIndex);
621 template <typename T, typename Walker>
622 struct generic_def_path_iterator
623 : public iterator_facade_base<generic_def_path_iterator<T, Walker>,
624 std::forward_iterator_tag, T *> {
625 generic_def_path_iterator() = default;
626 generic_def_path_iterator(Walker *W, ListIndex N) : W(W), N(N) {}
628 T &operator*() const { return curNode(); }
630 generic_def_path_iterator &operator++() {
631 N = curNode().Previous;
635 bool operator==(const generic_def_path_iterator &O) const {
636 if (N.hasValue() != O.N.hasValue())
638 return !N.hasValue() || *N == *O.N;
642 T &curNode() const { return W->Paths[*N]; }
645 Optional<ListIndex> N = None;
648 using def_path_iterator = generic_def_path_iterator<DefPath, ClobberWalker>;
649 using const_def_path_iterator =
650 generic_def_path_iterator<const DefPath, const ClobberWalker>;
652 iterator_range<def_path_iterator> def_path(ListIndex From) {
653 return make_range(def_path_iterator(this, From), def_path_iterator());
656 iterator_range<const_def_path_iterator> const_def_path(ListIndex From) const {
657 return make_range(const_def_path_iterator(this, From),
658 const_def_path_iterator());
662 /// The path that contains our result.
663 TerminatedPath PrimaryClobber;
664 /// The paths that we can legally cache back from, but that aren't
665 /// necessarily the result of the Phi optimization.
666 SmallVector<TerminatedPath, 4> OtherClobbers;
669 ListIndex defPathIndex(const DefPath &N) const {
670 // The assert looks nicer if we don't need to do &N
671 const DefPath *NP = &N;
672 assert(!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() &&
673 "Out of bounds DefPath!");
674 return NP - &Paths.front();
677 /// Try to optimize a phi as best as we can. Returns a SmallVector of Paths
678 /// that act as legal clobbers. Note that this won't return *all* clobbers.
680 /// Phi optimization algorithm tl;dr:
681 /// - Find the earliest def/phi, A, we can optimize to
682 /// - Find if all paths from the starting memory access ultimately reach A
683 /// - If not, optimization isn't possible.
684 /// - Otherwise, walk from A to another clobber or phi, A'.
685 /// - If A' is a def, we're done.
686 /// - If A' is a phi, try to optimize it.
688 /// A path is a series of {MemoryAccess, MemoryLocation} pairs. A path
689 /// terminates when a MemoryAccess that clobbers said MemoryLocation is found.
690 OptznResult tryOptimizePhi(MemoryPhi *Phi, MemoryAccess *Start,
691 const MemoryLocation &Loc) {
692 assert(Paths.empty() && VisitedPhis.empty() &&
693 "Reset the optimization state.");
695 Paths.emplace_back(Loc, Start, Phi, None);
696 // Stores how many "valid" optimization nodes we had prior to calling
697 // addSearches/getBlockingAccess. Necessary for caching if we had a blocker.
698 auto PriorPathsSize = Paths.size();
700 SmallVector<ListIndex, 16> PausedSearches;
701 SmallVector<ListIndex, 8> NewPaused;
702 SmallVector<TerminatedPath, 4> TerminatedPaths;
704 addSearches(Phi, PausedSearches, 0);
706 // Moves the TerminatedPath with the "most dominated" Clobber to the end of
708 auto MoveDominatedPathToEnd = [&](SmallVectorImpl<TerminatedPath> &Paths) {
709 assert(!Paths.empty() && "Need a path to move");
710 auto Dom = Paths.begin();
711 for (auto I = std::next(Dom), E = Paths.end(); I != E; ++I)
712 if (!MSSA.dominates(I->Clobber, Dom->Clobber))
714 auto Last = Paths.end() - 1;
716 std::iter_swap(Last, Dom);
719 MemoryPhi *Current = Phi;
721 assert(!MSSA.isLiveOnEntryDef(Current) &&
722 "liveOnEntry wasn't treated as a clobber?");
724 const auto *Target = getWalkTarget(Current);
725 // If a TerminatedPath doesn't dominate Target, then it wasn't a legal
726 // optimization for the prior phi.
727 assert(all_of(TerminatedPaths, [&](const TerminatedPath &P) {
728 return MSSA.dominates(P.Clobber, Target);
731 // FIXME: This is broken, because the Blocker may be reported to be
732 // liveOnEntry, and we'll happily wait for that to disappear (read: never)
733 // For the moment, this is fine, since we do nothing with blocker info.
734 if (Optional<TerminatedPath> Blocker = getBlockingAccess(
735 Target, PausedSearches, NewPaused, TerminatedPaths)) {
737 // Find the node we started at. We can't search based on N->Last, since
738 // we may have gone around a loop with a different MemoryLocation.
739 auto Iter = find_if(def_path(Blocker->LastNode), [&](const DefPath &N) {
740 return defPathIndex(N) < PriorPathsSize;
742 assert(Iter != def_path_iterator());
744 DefPath &CurNode = *Iter;
745 assert(CurNode.Last == Current);
748 // A. We can't reliably cache all of NewPaused back. Consider a case
749 // where we have two paths in NewPaused; one of which can't optimize
750 // above this phi, whereas the other can. If we cache the second path
751 // back, we'll end up with suboptimal cache entries. We can handle
752 // cases like this a bit better when we either try to find all
753 // clobbers that block phi optimization, or when our cache starts
754 // supporting unfinished searches.
755 // B. We can't reliably cache TerminatedPaths back here without doing
756 // extra checks; consider a case like:
762 // Where T is our target, C is a node with a clobber on it, D is a
763 // diamond (with a clobber *only* on the left or right node, N), and
764 // S is our start. Say we walk to D, through the node opposite N
765 // (read: ignoring the clobber), and see a cache entry in the top
766 // node of D. That cache entry gets put into TerminatedPaths. We then
767 // walk up to C (N is later in our worklist), find the clobber, and
768 // quit. If we append TerminatedPaths to OtherClobbers, we'll cache
769 // the bottom part of D to the cached clobber, ignoring the clobber
770 // in N. Again, this problem goes away if we start tracking all
771 // blockers for a given phi optimization.
772 TerminatedPath Result{CurNode.Last, defPathIndex(CurNode)};
776 // If there's nothing left to search, then all paths led to valid clobbers
777 // that we got from our cache; pick the nearest to the start, and allow
778 // the rest to be cached back.
779 if (NewPaused.empty()) {
780 MoveDominatedPathToEnd(TerminatedPaths);
781 TerminatedPath Result = TerminatedPaths.pop_back_val();
782 return {Result, std::move(TerminatedPaths)};
785 MemoryAccess *DefChainEnd = nullptr;
786 SmallVector<TerminatedPath, 4> Clobbers;
787 for (ListIndex Paused : NewPaused) {
788 UpwardsWalkResult WR = walkToPhiOrClobber(Paths[Paused]);
789 if (WR.IsKnownClobber)
790 Clobbers.push_back({WR.Result, Paused});
792 // Micro-opt: If we hit the end of the chain, save it.
793 DefChainEnd = WR.Result;
796 if (!TerminatedPaths.empty()) {
797 // If we couldn't find the dominating phi/liveOnEntry in the above loop,
800 for (auto *MA : def_chain(const_cast<MemoryAccess *>(Target)))
803 // If any of the terminated paths don't dominate the phi we'll try to
804 // optimize, we need to figure out what they are and quit.
805 const BasicBlock *ChainBB = DefChainEnd->getBlock();
806 for (const TerminatedPath &TP : TerminatedPaths) {
807 // Because we know that DefChainEnd is as "high" as we can go, we
808 // don't need local dominance checks; BB dominance is sufficient.
809 if (DT.dominates(ChainBB, TP.Clobber->getBlock()))
810 Clobbers.push_back(TP);
814 // If we have clobbers in the def chain, find the one closest to Current
816 if (!Clobbers.empty()) {
817 MoveDominatedPathToEnd(Clobbers);
818 TerminatedPath Result = Clobbers.pop_back_val();
819 return {Result, std::move(Clobbers)};
822 assert(all_of(NewPaused,
823 [&](ListIndex I) { return Paths[I].Last == DefChainEnd; }));
825 // Because liveOnEntry is a clobber, this must be a phi.
826 auto *DefChainPhi = cast<MemoryPhi>(DefChainEnd);
828 PriorPathsSize = Paths.size();
829 PausedSearches.clear();
830 for (ListIndex I : NewPaused)
831 addSearches(DefChainPhi, PausedSearches, I);
834 Current = DefChainPhi;
838 void verifyOptResult(const OptznResult &R) const {
839 assert(all_of(R.OtherClobbers, [&](const TerminatedPath &P) {
840 return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber);
844 void resetPhiOptznState() {
850 ClobberWalker(const MemorySSA &MSSA, AliasAnalysis &AA, DominatorTree &DT)
851 : MSSA(MSSA), AA(AA), DT(DT) {}
853 /// Finds the nearest clobber for the given query, optimizing phis if
855 MemoryAccess *findClobber(MemoryAccess *Start, UpwardsMemoryQuery &Q) {
858 MemoryAccess *Current = Start;
859 // This walker pretends uses don't exist. If we're handed one, silently grab
860 // its def. (This has the nice side-effect of ensuring we never cache uses)
861 if (auto *MU = dyn_cast<MemoryUse>(Start))
862 Current = MU->getDefiningAccess();
864 DefPath FirstDesc(Q.StartingLoc, Current, Current, None);
865 // Fast path for the overly-common case (no crazy phi optimization
867 UpwardsWalkResult WalkResult = walkToPhiOrClobber(FirstDesc);
868 MemoryAccess *Result;
869 if (WalkResult.IsKnownClobber) {
870 Result = WalkResult.Result;
871 Q.AR = WalkResult.AR;
873 OptznResult OptRes = tryOptimizePhi(cast<MemoryPhi>(FirstDesc.Last),
874 Current, Q.StartingLoc);
875 verifyOptResult(OptRes);
876 resetPhiOptznState();
877 Result = OptRes.PrimaryClobber.Clobber;
880 #ifdef EXPENSIVE_CHECKS
881 checkClobberSanity(Current, Result, Q.StartingLoc, MSSA, Q, AA);
886 void verify(const MemorySSA *MSSA) { assert(MSSA == &this->MSSA); }
889 struct RenamePassData {
891 DomTreeNode::const_iterator ChildIt;
892 MemoryAccess *IncomingVal;
894 RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It,
896 : DTN(D), ChildIt(It), IncomingVal(M) {}
898 void swap(RenamePassData &RHS) {
899 std::swap(DTN, RHS.DTN);
900 std::swap(ChildIt, RHS.ChildIt);
901 std::swap(IncomingVal, RHS.IncomingVal);
905 } // end anonymous namespace
909 /// A MemorySSAWalker that does AA walks to disambiguate accesses. It no
910 /// longer does caching on its own,
911 /// but the name has been retained for the moment.
912 class MemorySSA::CachingWalker final : public MemorySSAWalker {
913 ClobberWalker Walker;
915 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, UpwardsMemoryQuery &);
918 CachingWalker(MemorySSA *, AliasAnalysis *, DominatorTree *);
919 ~CachingWalker() override = default;
921 using MemorySSAWalker::getClobberingMemoryAccess;
923 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *) override;
924 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *,
925 const MemoryLocation &) override;
926 void invalidateInfo(MemoryAccess *) override;
928 void verify(const MemorySSA *MSSA) override {
929 MemorySSAWalker::verify(MSSA);
934 } // end namespace llvm
936 void MemorySSA::renameSuccessorPhis(BasicBlock *BB, MemoryAccess *IncomingVal,
937 bool RenameAllUses) {
938 // Pass through values to our successors
939 for (const BasicBlock *S : successors(BB)) {
940 auto It = PerBlockAccesses.find(S);
941 // Rename the phi nodes in our successor block
942 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
944 AccessList *Accesses = It->second.get();
945 auto *Phi = cast<MemoryPhi>(&Accesses->front());
947 int PhiIndex = Phi->getBasicBlockIndex(BB);
948 assert(PhiIndex != -1 && "Incomplete phi during partial rename");
949 Phi->setIncomingValue(PhiIndex, IncomingVal);
951 Phi->addIncoming(IncomingVal, BB);
955 /// Rename a single basic block into MemorySSA form.
956 /// Uses the standard SSA renaming algorithm.
957 /// \returns The new incoming value.
958 MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB, MemoryAccess *IncomingVal,
959 bool RenameAllUses) {
960 auto It = PerBlockAccesses.find(BB);
961 // Skip most processing if the list is empty.
962 if (It != PerBlockAccesses.end()) {
963 AccessList *Accesses = It->second.get();
964 for (MemoryAccess &L : *Accesses) {
965 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(&L)) {
966 if (MUD->getDefiningAccess() == nullptr || RenameAllUses)
967 MUD->setDefiningAccess(IncomingVal);
968 if (isa<MemoryDef>(&L))
978 /// This is the standard SSA renaming algorithm.
980 /// We walk the dominator tree in preorder, renaming accesses, and then filling
981 /// in phi nodes in our successors.
982 void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal,
983 SmallPtrSetImpl<BasicBlock *> &Visited,
984 bool SkipVisited, bool RenameAllUses) {
985 SmallVector<RenamePassData, 32> WorkStack;
986 // Skip everything if we already renamed this block and we are skipping.
987 // Note: You can't sink this into the if, because we need it to occur
988 // regardless of whether we skip blocks or not.
989 bool AlreadyVisited = !Visited.insert(Root->getBlock()).second;
990 if (SkipVisited && AlreadyVisited)
993 IncomingVal = renameBlock(Root->getBlock(), IncomingVal, RenameAllUses);
994 renameSuccessorPhis(Root->getBlock(), IncomingVal, RenameAllUses);
995 WorkStack.push_back({Root, Root->begin(), IncomingVal});
997 while (!WorkStack.empty()) {
998 DomTreeNode *Node = WorkStack.back().DTN;
999 DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt;
1000 IncomingVal = WorkStack.back().IncomingVal;
1002 if (ChildIt == Node->end()) {
1003 WorkStack.pop_back();
1005 DomTreeNode *Child = *ChildIt;
1006 ++WorkStack.back().ChildIt;
1007 BasicBlock *BB = Child->getBlock();
1008 // Note: You can't sink this into the if, because we need it to occur
1009 // regardless of whether we skip blocks or not.
1010 AlreadyVisited = !Visited.insert(BB).second;
1011 if (SkipVisited && AlreadyVisited) {
1012 // We already visited this during our renaming, which can happen when
1013 // being asked to rename multiple blocks. Figure out the incoming val,
1014 // which is the last def.
1015 // Incoming value can only change if there is a block def, and in that
1016 // case, it's the last block def in the list.
1017 if (auto *BlockDefs = getWritableBlockDefs(BB))
1018 IncomingVal = &*BlockDefs->rbegin();
1020 IncomingVal = renameBlock(BB, IncomingVal, RenameAllUses);
1021 renameSuccessorPhis(BB, IncomingVal, RenameAllUses);
1022 WorkStack.push_back({Child, Child->begin(), IncomingVal});
1027 /// This handles unreachable block accesses by deleting phi nodes in
1028 /// unreachable blocks, and marking all other unreachable MemoryAccess's as
1029 /// being uses of the live on entry definition.
1030 void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) {
1031 assert(!DT->isReachableFromEntry(BB) &&
1032 "Reachable block found while handling unreachable blocks");
1034 // Make sure phi nodes in our reachable successors end up with a
1035 // LiveOnEntryDef for our incoming edge, even though our block is forward
1036 // unreachable. We could just disconnect these blocks from the CFG fully,
1037 // but we do not right now.
1038 for (const BasicBlock *S : successors(BB)) {
1039 if (!DT->isReachableFromEntry(S))
1041 auto It = PerBlockAccesses.find(S);
1042 // Rename the phi nodes in our successor block
1043 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
1045 AccessList *Accesses = It->second.get();
1046 auto *Phi = cast<MemoryPhi>(&Accesses->front());
1047 Phi->addIncoming(LiveOnEntryDef.get(), BB);
1050 auto It = PerBlockAccesses.find(BB);
1051 if (It == PerBlockAccesses.end())
1054 auto &Accesses = It->second;
1055 for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) {
1056 auto Next = std::next(AI);
1057 // If we have a phi, just remove it. We are going to replace all
1058 // users with live on entry.
1059 if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(AI))
1060 UseOrDef->setDefiningAccess(LiveOnEntryDef.get());
1062 Accesses->erase(AI);
1067 MemorySSA::MemorySSA(Function &Func, AliasAnalysis *AA, DominatorTree *DT)
1068 : AA(AA), DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr),
1073 MemorySSA::~MemorySSA() {
1074 // Drop all our references
1075 for (const auto &Pair : PerBlockAccesses)
1076 for (MemoryAccess &MA : *Pair.second)
1077 MA.dropAllReferences();
1080 MemorySSA::AccessList *MemorySSA::getOrCreateAccessList(const BasicBlock *BB) {
1081 auto Res = PerBlockAccesses.insert(std::make_pair(BB, nullptr));
1084 Res.first->second = llvm::make_unique<AccessList>();
1085 return Res.first->second.get();
1088 MemorySSA::DefsList *MemorySSA::getOrCreateDefsList(const BasicBlock *BB) {
1089 auto Res = PerBlockDefs.insert(std::make_pair(BB, nullptr));
1092 Res.first->second = llvm::make_unique<DefsList>();
1093 return Res.first->second.get();
1098 /// This class is a batch walker of all MemoryUse's in the program, and points
1099 /// their defining access at the thing that actually clobbers them. Because it
1100 /// is a batch walker that touches everything, it does not operate like the
1101 /// other walkers. This walker is basically performing a top-down SSA renaming
1102 /// pass, where the version stack is used as the cache. This enables it to be
1103 /// significantly more time and memory efficient than using the regular walker,
1104 /// which is walking bottom-up.
1105 class MemorySSA::OptimizeUses {
1107 OptimizeUses(MemorySSA *MSSA, MemorySSAWalker *Walker, AliasAnalysis *AA,
1109 : MSSA(MSSA), Walker(Walker), AA(AA), DT(DT) {
1110 Walker = MSSA->getWalker();
1113 void optimizeUses();
1116 /// This represents where a given memorylocation is in the stack.
1117 struct MemlocStackInfo {
1118 // This essentially is keeping track of versions of the stack. Whenever
1119 // the stack changes due to pushes or pops, these versions increase.
1120 unsigned long StackEpoch;
1121 unsigned long PopEpoch;
1122 // This is the lower bound of places on the stack to check. It is equal to
1123 // the place the last stack walk ended.
1124 // Note: Correctness depends on this being initialized to 0, which densemap
1126 unsigned long LowerBound;
1127 const BasicBlock *LowerBoundBlock;
1128 // This is where the last walk for this memory location ended.
1129 unsigned long LastKill;
1131 Optional<AliasResult> AR;
1134 void optimizeUsesInBlock(const BasicBlock *, unsigned long &, unsigned long &,
1135 SmallVectorImpl<MemoryAccess *> &,
1136 DenseMap<MemoryLocOrCall, MemlocStackInfo> &);
1139 MemorySSAWalker *Walker;
1144 } // end namespace llvm
1146 /// Optimize the uses in a given block This is basically the SSA renaming
1147 /// algorithm, with one caveat: We are able to use a single stack for all
1148 /// MemoryUses. This is because the set of *possible* reaching MemoryDefs is
1149 /// the same for every MemoryUse. The *actual* clobbering MemoryDef is just
1150 /// going to be some position in that stack of possible ones.
1152 /// We track the stack positions that each MemoryLocation needs
1153 /// to check, and last ended at. This is because we only want to check the
1154 /// things that changed since last time. The same MemoryLocation should
1155 /// get clobbered by the same store (getModRefInfo does not use invariantness or
1156 /// things like this, and if they start, we can modify MemoryLocOrCall to
1157 /// include relevant data)
1158 void MemorySSA::OptimizeUses::optimizeUsesInBlock(
1159 const BasicBlock *BB, unsigned long &StackEpoch, unsigned long &PopEpoch,
1160 SmallVectorImpl<MemoryAccess *> &VersionStack,
1161 DenseMap<MemoryLocOrCall, MemlocStackInfo> &LocStackInfo) {
1163 /// If no accesses, nothing to do.
1164 MemorySSA::AccessList *Accesses = MSSA->getWritableBlockAccesses(BB);
1165 if (Accesses == nullptr)
1168 // Pop everything that doesn't dominate the current block off the stack,
1169 // increment the PopEpoch to account for this.
1172 !VersionStack.empty() &&
1173 "Version stack should have liveOnEntry sentinel dominating everything");
1174 BasicBlock *BackBlock = VersionStack.back()->getBlock();
1175 if (DT->dominates(BackBlock, BB))
1177 while (VersionStack.back()->getBlock() == BackBlock)
1178 VersionStack.pop_back();
1182 for (MemoryAccess &MA : *Accesses) {
1183 auto *MU = dyn_cast<MemoryUse>(&MA);
1185 VersionStack.push_back(&MA);
1190 if (isUseTriviallyOptimizableToLiveOnEntry(*AA, MU->getMemoryInst())) {
1191 MU->setDefiningAccess(MSSA->getLiveOnEntryDef(), true, None);
1195 MemoryLocOrCall UseMLOC(MU);
1196 auto &LocInfo = LocStackInfo[UseMLOC];
1197 // If the pop epoch changed, it means we've removed stuff from top of
1198 // stack due to changing blocks. We may have to reset the lower bound or
1200 if (LocInfo.PopEpoch != PopEpoch) {
1201 LocInfo.PopEpoch = PopEpoch;
1202 LocInfo.StackEpoch = StackEpoch;
1203 // If the lower bound was in something that no longer dominates us, we
1204 // have to reset it.
1205 // We can't simply track stack size, because the stack may have had
1206 // pushes/pops in the meantime.
1207 // XXX: This is non-optimal, but only is slower cases with heavily
1208 // branching dominator trees. To get the optimal number of queries would
1209 // be to make lowerbound and lastkill a per-loc stack, and pop it until
1210 // the top of that stack dominates us. This does not seem worth it ATM.
1211 // A much cheaper optimization would be to always explore the deepest
1212 // branch of the dominator tree first. This will guarantee this resets on
1213 // the smallest set of blocks.
1214 if (LocInfo.LowerBoundBlock && LocInfo.LowerBoundBlock != BB &&
1215 !DT->dominates(LocInfo.LowerBoundBlock, BB)) {
1216 // Reset the lower bound of things to check.
1217 // TODO: Some day we should be able to reset to last kill, rather than
1219 LocInfo.LowerBound = 0;
1220 LocInfo.LowerBoundBlock = VersionStack[0]->getBlock();
1221 LocInfo.LastKillValid = false;
1223 } else if (LocInfo.StackEpoch != StackEpoch) {
1224 // If all that has changed is the StackEpoch, we only have to check the
1225 // new things on the stack, because we've checked everything before. In
1226 // this case, the lower bound of things to check remains the same.
1227 LocInfo.PopEpoch = PopEpoch;
1228 LocInfo.StackEpoch = StackEpoch;
1230 if (!LocInfo.LastKillValid) {
1231 LocInfo.LastKill = VersionStack.size() - 1;
1232 LocInfo.LastKillValid = true;
1233 LocInfo.AR = MayAlias;
1236 // At this point, we should have corrected last kill and LowerBound to be
1238 assert(LocInfo.LowerBound < VersionStack.size() &&
1239 "Lower bound out of range");
1240 assert(LocInfo.LastKill < VersionStack.size() &&
1241 "Last kill info out of range");
1242 // In any case, the new upper bound is the top of the stack.
1243 unsigned long UpperBound = VersionStack.size() - 1;
1245 if (UpperBound - LocInfo.LowerBound > MaxCheckLimit) {
1246 DEBUG(dbgs() << "MemorySSA skipping optimization of " << *MU << " ("
1247 << *(MU->getMemoryInst()) << ")"
1248 << " because there are " << UpperBound - LocInfo.LowerBound
1249 << " stores to disambiguate\n");
1250 // Because we did not walk, LastKill is no longer valid, as this may
1251 // have been a kill.
1252 LocInfo.LastKillValid = false;
1255 bool FoundClobberResult = false;
1256 while (UpperBound > LocInfo.LowerBound) {
1257 if (isa<MemoryPhi>(VersionStack[UpperBound])) {
1258 // For phis, use the walker, see where we ended up, go there
1259 Instruction *UseInst = MU->getMemoryInst();
1260 MemoryAccess *Result = Walker->getClobberingMemoryAccess(UseInst);
1261 // We are guaranteed to find it or something is wrong
1262 while (VersionStack[UpperBound] != Result) {
1263 assert(UpperBound != 0);
1266 FoundClobberResult = true;
1270 MemoryDef *MD = cast<MemoryDef>(VersionStack[UpperBound]);
1271 // If the lifetime of the pointer ends at this instruction, it's live on
1273 if (!UseMLOC.IsCall && lifetimeEndsAt(MD, UseMLOC.getLoc(), *AA)) {
1274 // Reset UpperBound to liveOnEntryDef's place in the stack
1276 FoundClobberResult = true;
1277 LocInfo.AR = MustAlias;
1280 ClobberAlias CA = instructionClobbersQuery(MD, MU, UseMLOC, *AA);
1282 FoundClobberResult = true;
1289 // Note: Phis always have AliasResult AR set to MayAlias ATM.
1291 // At the end of this loop, UpperBound is either a clobber, or lower bound
1292 // PHI walking may cause it to be < LowerBound, and in fact, < LastKill.
1293 if (FoundClobberResult || UpperBound < LocInfo.LastKill) {
1294 // We were last killed now by where we got to
1295 if (MSSA->isLiveOnEntryDef(VersionStack[UpperBound]))
1297 MU->setDefiningAccess(VersionStack[UpperBound], true, LocInfo.AR);
1298 LocInfo.LastKill = UpperBound;
1300 // Otherwise, we checked all the new ones, and now we know we can get to
1302 MU->setDefiningAccess(VersionStack[LocInfo.LastKill], true, LocInfo.AR);
1304 LocInfo.LowerBound = VersionStack.size() - 1;
1305 LocInfo.LowerBoundBlock = BB;
1309 /// Optimize uses to point to their actual clobbering definitions.
1310 void MemorySSA::OptimizeUses::optimizeUses() {
1311 SmallVector<MemoryAccess *, 16> VersionStack;
1312 DenseMap<MemoryLocOrCall, MemlocStackInfo> LocStackInfo;
1313 VersionStack.push_back(MSSA->getLiveOnEntryDef());
1315 unsigned long StackEpoch = 1;
1316 unsigned long PopEpoch = 1;
1317 // We perform a non-recursive top-down dominator tree walk.
1318 for (const auto *DomNode : depth_first(DT->getRootNode()))
1319 optimizeUsesInBlock(DomNode->getBlock(), StackEpoch, PopEpoch, VersionStack,
1323 void MemorySSA::placePHINodes(
1324 const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks,
1325 const DenseMap<const BasicBlock *, unsigned int> &BBNumbers) {
1326 // Determine where our MemoryPhi's should go
1327 ForwardIDFCalculator IDFs(*DT);
1328 IDFs.setDefiningBlocks(DefiningBlocks);
1329 SmallVector<BasicBlock *, 32> IDFBlocks;
1330 IDFs.calculate(IDFBlocks);
1332 llvm::sort(IDFBlocks.begin(), IDFBlocks.end(),
1333 [&BBNumbers](const BasicBlock *A, const BasicBlock *B) {
1334 return BBNumbers.lookup(A) < BBNumbers.lookup(B);
1337 // Now place MemoryPhi nodes.
1338 for (auto &BB : IDFBlocks)
1339 createMemoryPhi(BB);
1342 void MemorySSA::buildMemorySSA() {
1343 // We create an access to represent "live on entry", for things like
1344 // arguments or users of globals, where the memory they use is defined before
1345 // the beginning of the function. We do not actually insert it into the IR.
1346 // We do not define a live on exit for the immediate uses, and thus our
1347 // semantics do *not* imply that something with no immediate uses can simply
1349 BasicBlock &StartingPoint = F.getEntryBlock();
1350 LiveOnEntryDef.reset(new MemoryDef(F.getContext(), nullptr, nullptr,
1351 &StartingPoint, NextID++));
1352 DenseMap<const BasicBlock *, unsigned int> BBNumbers;
1353 unsigned NextBBNum = 0;
1355 // We maintain lists of memory accesses per-block, trading memory for time. We
1356 // could just look up the memory access for every possible instruction in the
1358 SmallPtrSet<BasicBlock *, 32> DefiningBlocks;
1359 // Go through each block, figure out where defs occur, and chain together all
1361 for (BasicBlock &B : F) {
1362 BBNumbers[&B] = NextBBNum++;
1363 bool InsertIntoDef = false;
1364 AccessList *Accesses = nullptr;
1365 DefsList *Defs = nullptr;
1366 for (Instruction &I : B) {
1367 MemoryUseOrDef *MUD = createNewAccess(&I);
1372 Accesses = getOrCreateAccessList(&B);
1373 Accesses->push_back(MUD);
1374 if (isa<MemoryDef>(MUD)) {
1375 InsertIntoDef = true;
1377 Defs = getOrCreateDefsList(&B);
1378 Defs->push_back(*MUD);
1382 DefiningBlocks.insert(&B);
1384 placePHINodes(DefiningBlocks, BBNumbers);
1386 // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get
1387 // filled in with all blocks.
1388 SmallPtrSet<BasicBlock *, 16> Visited;
1389 renamePass(DT->getRootNode(), LiveOnEntryDef.get(), Visited);
1391 CachingWalker *Walker = getWalkerImpl();
1393 OptimizeUses(this, Walker, AA, DT).optimizeUses();
1395 // Mark the uses in unreachable blocks as live on entry, so that they go
1398 if (!Visited.count(&BB))
1399 markUnreachableAsLiveOnEntry(&BB);
1402 MemorySSAWalker *MemorySSA::getWalker() { return getWalkerImpl(); }
1404 MemorySSA::CachingWalker *MemorySSA::getWalkerImpl() {
1406 return Walker.get();
1408 Walker = llvm::make_unique<CachingWalker>(this, AA, DT);
1409 return Walker.get();
1412 // This is a helper function used by the creation routines. It places NewAccess
1413 // into the access and defs lists for a given basic block, at the given
1415 void MemorySSA::insertIntoListsForBlock(MemoryAccess *NewAccess,
1416 const BasicBlock *BB,
1417 InsertionPlace Point) {
1418 auto *Accesses = getOrCreateAccessList(BB);
1419 if (Point == Beginning) {
1420 // If it's a phi node, it goes first, otherwise, it goes after any phi
1422 if (isa<MemoryPhi>(NewAccess)) {
1423 Accesses->push_front(NewAccess);
1424 auto *Defs = getOrCreateDefsList(BB);
1425 Defs->push_front(*NewAccess);
1427 auto AI = find_if_not(
1428 *Accesses, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
1429 Accesses->insert(AI, NewAccess);
1430 if (!isa<MemoryUse>(NewAccess)) {
1431 auto *Defs = getOrCreateDefsList(BB);
1432 auto DI = find_if_not(
1433 *Defs, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
1434 Defs->insert(DI, *NewAccess);
1438 Accesses->push_back(NewAccess);
1439 if (!isa<MemoryUse>(NewAccess)) {
1440 auto *Defs = getOrCreateDefsList(BB);
1441 Defs->push_back(*NewAccess);
1444 BlockNumberingValid.erase(BB);
1447 void MemorySSA::insertIntoListsBefore(MemoryAccess *What, const BasicBlock *BB,
1448 AccessList::iterator InsertPt) {
1449 auto *Accesses = getWritableBlockAccesses(BB);
1450 bool WasEnd = InsertPt == Accesses->end();
1451 Accesses->insert(AccessList::iterator(InsertPt), What);
1452 if (!isa<MemoryUse>(What)) {
1453 auto *Defs = getOrCreateDefsList(BB);
1454 // If we got asked to insert at the end, we have an easy job, just shove it
1455 // at the end. If we got asked to insert before an existing def, we also get
1456 // an iterator. If we got asked to insert before a use, we have to hunt for
1459 Defs->push_back(*What);
1460 } else if (isa<MemoryDef>(InsertPt)) {
1461 Defs->insert(InsertPt->getDefsIterator(), *What);
1463 while (InsertPt != Accesses->end() && !isa<MemoryDef>(InsertPt))
1465 // Either we found a def, or we are inserting at the end
1466 if (InsertPt == Accesses->end())
1467 Defs->push_back(*What);
1469 Defs->insert(InsertPt->getDefsIterator(), *What);
1472 BlockNumberingValid.erase(BB);
1475 // Move What before Where in the IR. The end result is that What will belong to
1476 // the right lists and have the right Block set, but will not otherwise be
1477 // correct. It will not have the right defining access, and if it is a def,
1478 // things below it will not properly be updated.
1479 void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB,
1480 AccessList::iterator Where) {
1481 // Keep it in the lookup tables, remove from the lists
1482 removeFromLists(What, false);
1484 insertIntoListsBefore(What, BB, Where);
1487 void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB,
1488 InsertionPlace Point) {
1489 removeFromLists(What, false);
1491 insertIntoListsForBlock(What, BB, Point);
1494 MemoryPhi *MemorySSA::createMemoryPhi(BasicBlock *BB) {
1495 assert(!getMemoryAccess(BB) && "MemoryPhi already exists for this BB");
1496 MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++);
1497 // Phi's always are placed at the front of the block.
1498 insertIntoListsForBlock(Phi, BB, Beginning);
1499 ValueToMemoryAccess[BB] = Phi;
1503 MemoryUseOrDef *MemorySSA::createDefinedAccess(Instruction *I,
1504 MemoryAccess *Definition) {
1505 assert(!isa<PHINode>(I) && "Cannot create a defined access for a PHI");
1506 MemoryUseOrDef *NewAccess = createNewAccess(I);
1508 NewAccess != nullptr &&
1509 "Tried to create a memory access for a non-memory touching instruction");
1510 NewAccess->setDefiningAccess(Definition);
1514 // Return true if the instruction has ordering constraints.
1515 // Note specifically that this only considers stores and loads
1516 // because others are still considered ModRef by getModRefInfo.
1517 static inline bool isOrdered(const Instruction *I) {
1518 if (auto *SI = dyn_cast<StoreInst>(I)) {
1519 if (!SI->isUnordered())
1521 } else if (auto *LI = dyn_cast<LoadInst>(I)) {
1522 if (!LI->isUnordered())
1528 /// Helper function to create new memory accesses
1529 MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I) {
1530 // The assume intrinsic has a control dependency which we model by claiming
1531 // that it writes arbitrarily. Ignore that fake memory dependency here.
1532 // FIXME: Replace this special casing with a more accurate modelling of
1533 // assume's control dependency.
1534 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1535 if (II->getIntrinsicID() == Intrinsic::assume)
1538 // Find out what affect this instruction has on memory.
1539 ModRefInfo ModRef = AA->getModRefInfo(I, None);
1540 // The isOrdered check is used to ensure that volatiles end up as defs
1541 // (atomics end up as ModRef right now anyway). Until we separate the
1542 // ordering chain from the memory chain, this enables people to see at least
1543 // some relative ordering to volatiles. Note that getClobberingMemoryAccess
1544 // will still give an answer that bypasses other volatile loads. TODO:
1545 // Separate memory aliasing and ordering into two different chains so that we
1546 // can precisely represent both "what memory will this read/write/is clobbered
1547 // by" and "what instructions can I move this past".
1548 bool Def = isModSet(ModRef) || isOrdered(I);
1549 bool Use = isRefSet(ModRef);
1551 // It's possible for an instruction to not modify memory at all. During
1552 // construction, we ignore them.
1556 MemoryUseOrDef *MUD;
1558 MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++);
1560 MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent());
1561 ValueToMemoryAccess[I] = MUD;
1565 /// Returns true if \p Replacer dominates \p Replacee .
1566 bool MemorySSA::dominatesUse(const MemoryAccess *Replacer,
1567 const MemoryAccess *Replacee) const {
1568 if (isa<MemoryUseOrDef>(Replacee))
1569 return DT->dominates(Replacer->getBlock(), Replacee->getBlock());
1570 const auto *MP = cast<MemoryPhi>(Replacee);
1571 // For a phi node, the use occurs in the predecessor block of the phi node.
1572 // Since we may occur multiple times in the phi node, we have to check each
1573 // operand to ensure Replacer dominates each operand where Replacee occurs.
1574 for (const Use &Arg : MP->operands()) {
1575 if (Arg.get() != Replacee &&
1576 !DT->dominates(Replacer->getBlock(), MP->getIncomingBlock(Arg)))
1582 /// Properly remove \p MA from all of MemorySSA's lookup tables.
1583 void MemorySSA::removeFromLookups(MemoryAccess *MA) {
1584 assert(MA->use_empty() &&
1585 "Trying to remove memory access that still has uses");
1586 BlockNumbering.erase(MA);
1587 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA))
1588 MUD->setDefiningAccess(nullptr);
1589 // Invalidate our walker's cache if necessary
1590 if (!isa<MemoryUse>(MA))
1591 Walker->invalidateInfo(MA);
1592 // The call below to erase will destroy MA, so we can't change the order we
1593 // are doing things here
1595 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA)) {
1596 MemoryInst = MUD->getMemoryInst();
1598 MemoryInst = MA->getBlock();
1600 auto VMA = ValueToMemoryAccess.find(MemoryInst);
1601 if (VMA->second == MA)
1602 ValueToMemoryAccess.erase(VMA);
1605 /// Properly remove \p MA from all of MemorySSA's lists.
1607 /// Because of the way the intrusive list and use lists work, it is important to
1608 /// do removal in the right order.
1609 /// ShouldDelete defaults to true, and will cause the memory access to also be
1610 /// deleted, not just removed.
1611 void MemorySSA::removeFromLists(MemoryAccess *MA, bool ShouldDelete) {
1612 // The access list owns the reference, so we erase it from the non-owning list
1614 if (!isa<MemoryUse>(MA)) {
1615 auto DefsIt = PerBlockDefs.find(MA->getBlock());
1616 std::unique_ptr<DefsList> &Defs = DefsIt->second;
1619 PerBlockDefs.erase(DefsIt);
1622 // The erase call here will delete it. If we don't want it deleted, we call
1624 auto AccessIt = PerBlockAccesses.find(MA->getBlock());
1625 std::unique_ptr<AccessList> &Accesses = AccessIt->second;
1627 Accesses->erase(MA);
1629 Accesses->remove(MA);
1631 if (Accesses->empty())
1632 PerBlockAccesses.erase(AccessIt);
1635 void MemorySSA::print(raw_ostream &OS) const {
1636 MemorySSAAnnotatedWriter Writer(this);
1637 F.print(OS, &Writer);
1640 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1641 LLVM_DUMP_METHOD void MemorySSA::dump() const { print(dbgs()); }
1644 void MemorySSA::verifyMemorySSA() const {
1646 verifyDomination(F);
1648 Walker->verify(this);
1651 /// Verify that the order and existence of MemoryAccesses matches the
1652 /// order and existence of memory affecting instructions.
1653 void MemorySSA::verifyOrdering(Function &F) const {
1654 // Walk all the blocks, comparing what the lookups think and what the access
1655 // lists think, as well as the order in the blocks vs the order in the access
1657 SmallVector<MemoryAccess *, 32> ActualAccesses;
1658 SmallVector<MemoryAccess *, 32> ActualDefs;
1659 for (BasicBlock &B : F) {
1660 const AccessList *AL = getBlockAccesses(&B);
1661 const auto *DL = getBlockDefs(&B);
1662 MemoryAccess *Phi = getMemoryAccess(&B);
1664 ActualAccesses.push_back(Phi);
1665 ActualDefs.push_back(Phi);
1668 for (Instruction &I : B) {
1669 MemoryAccess *MA = getMemoryAccess(&I);
1670 assert((!MA || (AL && (isa<MemoryUse>(MA) || DL))) &&
1671 "We have memory affecting instructions "
1672 "in this block but they are not in the "
1673 "access list or defs list");
1675 ActualAccesses.push_back(MA);
1676 if (isa<MemoryDef>(MA))
1677 ActualDefs.push_back(MA);
1680 // Either we hit the assert, really have no accesses, or we have both
1681 // accesses and an access list.
1685 assert(AL->size() == ActualAccesses.size() &&
1686 "We don't have the same number of accesses in the block as on the "
1688 assert((DL || ActualDefs.size() == 0) &&
1689 "Either we should have a defs list, or we should have no defs");
1690 assert((!DL || DL->size() == ActualDefs.size()) &&
1691 "We don't have the same number of defs in the block as on the "
1693 auto ALI = AL->begin();
1694 auto AAI = ActualAccesses.begin();
1695 while (ALI != AL->end() && AAI != ActualAccesses.end()) {
1696 assert(&*ALI == *AAI && "Not the same accesses in the same order");
1700 ActualAccesses.clear();
1702 auto DLI = DL->begin();
1703 auto ADI = ActualDefs.begin();
1704 while (DLI != DL->end() && ADI != ActualDefs.end()) {
1705 assert(&*DLI == *ADI && "Not the same defs in the same order");
1714 /// Verify the domination properties of MemorySSA by checking that each
1715 /// definition dominates all of its uses.
1716 void MemorySSA::verifyDomination(Function &F) const {
1718 for (BasicBlock &B : F) {
1719 // Phi nodes are attached to basic blocks
1720 if (MemoryPhi *MP = getMemoryAccess(&B))
1721 for (const Use &U : MP->uses())
1722 assert(dominates(MP, U) && "Memory PHI does not dominate it's uses");
1724 for (Instruction &I : B) {
1725 MemoryAccess *MD = dyn_cast_or_null<MemoryDef>(getMemoryAccess(&I));
1729 for (const Use &U : MD->uses())
1730 assert(dominates(MD, U) && "Memory Def does not dominate it's uses");
1736 /// Verify the def-use lists in MemorySSA, by verifying that \p Use
1737 /// appears in the use list of \p Def.
1738 void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const {
1740 // The live on entry use may cause us to get a NULL def here
1742 assert(isLiveOnEntryDef(Use) &&
1743 "Null def but use not point to live on entry def");
1745 assert(is_contained(Def->users(), Use) &&
1746 "Did not find use in def's use list");
1750 /// Verify the immediate use information, by walking all the memory
1751 /// accesses and verifying that, for each use, it appears in the
1752 /// appropriate def's use list
1753 void MemorySSA::verifyDefUses(Function &F) const {
1754 for (BasicBlock &B : F) {
1755 // Phi nodes are attached to basic blocks
1756 if (MemoryPhi *Phi = getMemoryAccess(&B)) {
1757 assert(Phi->getNumOperands() == static_cast<unsigned>(std::distance(
1758 pred_begin(&B), pred_end(&B))) &&
1759 "Incomplete MemoryPhi Node");
1760 for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I)
1761 verifyUseInDefs(Phi->getIncomingValue(I), Phi);
1764 for (Instruction &I : B) {
1765 if (MemoryUseOrDef *MA = getMemoryAccess(&I)) {
1766 verifyUseInDefs(MA->getDefiningAccess(), MA);
1772 MemoryUseOrDef *MemorySSA::getMemoryAccess(const Instruction *I) const {
1773 return cast_or_null<MemoryUseOrDef>(ValueToMemoryAccess.lookup(I));
1776 MemoryPhi *MemorySSA::getMemoryAccess(const BasicBlock *BB) const {
1777 return cast_or_null<MemoryPhi>(ValueToMemoryAccess.lookup(cast<Value>(BB)));
1780 /// Perform a local numbering on blocks so that instruction ordering can be
1781 /// determined in constant time.
1782 /// TODO: We currently just number in order. If we numbered by N, we could
1783 /// allow at least N-1 sequences of insertBefore or insertAfter (and at least
1784 /// log2(N) sequences of mixed before and after) without needing to invalidate
1786 void MemorySSA::renumberBlock(const BasicBlock *B) const {
1787 // The pre-increment ensures the numbers really start at 1.
1788 unsigned long CurrentNumber = 0;
1789 const AccessList *AL = getBlockAccesses(B);
1790 assert(AL != nullptr && "Asking to renumber an empty block");
1791 for (const auto &I : *AL)
1792 BlockNumbering[&I] = ++CurrentNumber;
1793 BlockNumberingValid.insert(B);
1796 /// Determine, for two memory accesses in the same block,
1797 /// whether \p Dominator dominates \p Dominatee.
1798 /// \returns True if \p Dominator dominates \p Dominatee.
1799 bool MemorySSA::locallyDominates(const MemoryAccess *Dominator,
1800 const MemoryAccess *Dominatee) const {
1801 const BasicBlock *DominatorBlock = Dominator->getBlock();
1803 assert((DominatorBlock == Dominatee->getBlock()) &&
1804 "Asking for local domination when accesses are in different blocks!");
1805 // A node dominates itself.
1806 if (Dominatee == Dominator)
1809 // When Dominatee is defined on function entry, it is not dominated by another
1811 if (isLiveOnEntryDef(Dominatee))
1814 // When Dominator is defined on function entry, it dominates the other memory
1816 if (isLiveOnEntryDef(Dominator))
1819 if (!BlockNumberingValid.count(DominatorBlock))
1820 renumberBlock(DominatorBlock);
1822 unsigned long DominatorNum = BlockNumbering.lookup(Dominator);
1823 // All numbers start with 1
1824 assert(DominatorNum != 0 && "Block was not numbered properly");
1825 unsigned long DominateeNum = BlockNumbering.lookup(Dominatee);
1826 assert(DominateeNum != 0 && "Block was not numbered properly");
1827 return DominatorNum < DominateeNum;
1830 bool MemorySSA::dominates(const MemoryAccess *Dominator,
1831 const MemoryAccess *Dominatee) const {
1832 if (Dominator == Dominatee)
1835 if (isLiveOnEntryDef(Dominatee))
1838 if (Dominator->getBlock() != Dominatee->getBlock())
1839 return DT->dominates(Dominator->getBlock(), Dominatee->getBlock());
1840 return locallyDominates(Dominator, Dominatee);
1843 bool MemorySSA::dominates(const MemoryAccess *Dominator,
1844 const Use &Dominatee) const {
1845 if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Dominatee.getUser())) {
1846 BasicBlock *UseBB = MP->getIncomingBlock(Dominatee);
1847 // The def must dominate the incoming block of the phi.
1848 if (UseBB != Dominator->getBlock())
1849 return DT->dominates(Dominator->getBlock(), UseBB);
1850 // If the UseBB and the DefBB are the same, compare locally.
1851 return locallyDominates(Dominator, cast<MemoryAccess>(Dominatee));
1853 // If it's not a PHI node use, the normal dominates can already handle it.
1854 return dominates(Dominator, cast<MemoryAccess>(Dominatee.getUser()));
1857 const static char LiveOnEntryStr[] = "liveOnEntry";
1859 void MemoryAccess::print(raw_ostream &OS) const {
1860 switch (getValueID()) {
1861 case MemoryPhiVal: return static_cast<const MemoryPhi *>(this)->print(OS);
1862 case MemoryDefVal: return static_cast<const MemoryDef *>(this)->print(OS);
1863 case MemoryUseVal: return static_cast<const MemoryUse *>(this)->print(OS);
1865 llvm_unreachable("invalid value id");
1868 void MemoryDef::print(raw_ostream &OS) const {
1869 MemoryAccess *UO = getDefiningAccess();
1871 OS << getID() << " = MemoryDef(";
1872 if (UO && UO->getID())
1875 OS << LiveOnEntryStr;
1879 void MemoryPhi::print(raw_ostream &OS) const {
1881 OS << getID() << " = MemoryPhi(";
1882 for (const auto &Op : operands()) {
1883 BasicBlock *BB = getIncomingBlock(Op);
1884 MemoryAccess *MA = cast<MemoryAccess>(Op);
1892 OS << BB->getName();
1894 BB->printAsOperand(OS, false);
1896 if (unsigned ID = MA->getID())
1899 OS << LiveOnEntryStr;
1905 void MemoryUse::print(raw_ostream &OS) const {
1906 MemoryAccess *UO = getDefiningAccess();
1908 if (UO && UO->getID())
1911 OS << LiveOnEntryStr;
1915 void MemoryAccess::dump() const {
1916 // Cannot completely remove virtual function even in release mode.
1917 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1923 char MemorySSAPrinterLegacyPass::ID = 0;
1925 MemorySSAPrinterLegacyPass::MemorySSAPrinterLegacyPass() : FunctionPass(ID) {
1926 initializeMemorySSAPrinterLegacyPassPass(*PassRegistry::getPassRegistry());
1929 void MemorySSAPrinterLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
1930 AU.setPreservesAll();
1931 AU.addRequired<MemorySSAWrapperPass>();
1934 bool MemorySSAPrinterLegacyPass::runOnFunction(Function &F) {
1935 auto &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA();
1937 if (VerifyMemorySSA)
1938 MSSA.verifyMemorySSA();
1942 AnalysisKey MemorySSAAnalysis::Key;
1944 MemorySSAAnalysis::Result MemorySSAAnalysis::run(Function &F,
1945 FunctionAnalysisManager &AM) {
1946 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1947 auto &AA = AM.getResult<AAManager>(F);
1948 return MemorySSAAnalysis::Result(llvm::make_unique<MemorySSA>(F, &AA, &DT));
1951 PreservedAnalyses MemorySSAPrinterPass::run(Function &F,
1952 FunctionAnalysisManager &AM) {
1953 OS << "MemorySSA for function: " << F.getName() << "\n";
1954 AM.getResult<MemorySSAAnalysis>(F).getMSSA().print(OS);
1956 return PreservedAnalyses::all();
1959 PreservedAnalyses MemorySSAVerifierPass::run(Function &F,
1960 FunctionAnalysisManager &AM) {
1961 AM.getResult<MemorySSAAnalysis>(F).getMSSA().verifyMemorySSA();
1963 return PreservedAnalyses::all();
1966 char MemorySSAWrapperPass::ID = 0;
1968 MemorySSAWrapperPass::MemorySSAWrapperPass() : FunctionPass(ID) {
1969 initializeMemorySSAWrapperPassPass(*PassRegistry::getPassRegistry());
1972 void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); }
1974 void MemorySSAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1975 AU.setPreservesAll();
1976 AU.addRequiredTransitive<DominatorTreeWrapperPass>();
1977 AU.addRequiredTransitive<AAResultsWrapperPass>();
1980 bool MemorySSAWrapperPass::runOnFunction(Function &F) {
1981 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1982 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
1983 MSSA.reset(new MemorySSA(F, &AA, &DT));
1987 void MemorySSAWrapperPass::verifyAnalysis() const { MSSA->verifyMemorySSA(); }
1989 void MemorySSAWrapperPass::print(raw_ostream &OS, const Module *M) const {
1993 MemorySSAWalker::MemorySSAWalker(MemorySSA *M) : MSSA(M) {}
1995 MemorySSA::CachingWalker::CachingWalker(MemorySSA *M, AliasAnalysis *A,
1997 : MemorySSAWalker(M), Walker(*M, *A, *D) {}
1999 void MemorySSA::CachingWalker::invalidateInfo(MemoryAccess *MA) {
2000 if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
2001 MUD->resetOptimized();
2004 /// Walk the use-def chains starting at \p MA and find
2005 /// the MemoryAccess that actually clobbers Loc.
2007 /// \returns our clobbering memory access
2008 MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
2009 MemoryAccess *StartingAccess, UpwardsMemoryQuery &Q) {
2010 return Walker.findClobber(StartingAccess, Q);
2013 MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
2014 MemoryAccess *StartingAccess, const MemoryLocation &Loc) {
2015 if (isa<MemoryPhi>(StartingAccess))
2016 return StartingAccess;
2018 auto *StartingUseOrDef = cast<MemoryUseOrDef>(StartingAccess);
2019 if (MSSA->isLiveOnEntryDef(StartingUseOrDef))
2020 return StartingUseOrDef;
2022 Instruction *I = StartingUseOrDef->getMemoryInst();
2024 // Conservatively, fences are always clobbers, so don't perform the walk if we
2026 if (!ImmutableCallSite(I) && I->isFenceLike())
2027 return StartingUseOrDef;
2029 UpwardsMemoryQuery Q;
2030 Q.OriginalAccess = StartingUseOrDef;
2031 Q.StartingLoc = Loc;
2035 // Unlike the other function, do not walk to the def of a def, because we are
2036 // handed something we already believe is the clobbering access.
2037 MemoryAccess *DefiningAccess = isa<MemoryUse>(StartingUseOrDef)
2038 ? StartingUseOrDef->getDefiningAccess()
2041 MemoryAccess *Clobber = getClobberingMemoryAccess(DefiningAccess, Q);
2042 DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
2043 DEBUG(dbgs() << *StartingUseOrDef << "\n");
2044 DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
2045 DEBUG(dbgs() << *Clobber << "\n");
2050 MemorySSA::CachingWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
2051 auto *StartingAccess = dyn_cast<MemoryUseOrDef>(MA);
2052 // If this is a MemoryPhi, we can't do anything.
2053 if (!StartingAccess)
2056 // If this is an already optimized use or def, return the optimized result.
2057 // Note: Currently, we store the optimized def result in a separate field,
2058 // since we can't use the defining access.
2059 if (StartingAccess->isOptimized())
2060 return StartingAccess->getOptimized();
2062 const Instruction *I = StartingAccess->getMemoryInst();
2063 UpwardsMemoryQuery Q(I, StartingAccess);
2064 // We can't sanely do anything with a fence, since they conservatively clobber
2065 // all memory, and have no locations to get pointers from to try to
2067 if (!Q.IsCall && I->isFenceLike())
2068 return StartingAccess;
2070 if (isUseTriviallyOptimizableToLiveOnEntry(*MSSA->AA, I)) {
2071 MemoryAccess *LiveOnEntry = MSSA->getLiveOnEntryDef();
2072 StartingAccess->setOptimized(LiveOnEntry);
2073 StartingAccess->setOptimizedAccessType(None);
2077 // Start with the thing we already think clobbers this location
2078 MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess();
2080 // At this point, DefiningAccess may be the live on entry def.
2081 // If it is, we will not get a better result.
2082 if (MSSA->isLiveOnEntryDef(DefiningAccess)) {
2083 StartingAccess->setOptimized(DefiningAccess);
2084 StartingAccess->setOptimizedAccessType(None);
2085 return DefiningAccess;
2088 MemoryAccess *Result = getClobberingMemoryAccess(DefiningAccess, Q);
2089 DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
2090 DEBUG(dbgs() << *DefiningAccess << "\n");
2091 DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
2092 DEBUG(dbgs() << *Result << "\n");
2094 StartingAccess->setOptimized(Result);
2095 if (MSSA->isLiveOnEntryDef(Result))
2096 StartingAccess->setOptimizedAccessType(None);
2097 else if (Q.AR == MustAlias)
2098 StartingAccess->setOptimizedAccessType(MustAlias);
2104 DoNothingMemorySSAWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
2105 if (auto *Use = dyn_cast<MemoryUseOrDef>(MA))
2106 return Use->getDefiningAccess();
2110 MemoryAccess *DoNothingMemorySSAWalker::getClobberingMemoryAccess(
2111 MemoryAccess *StartingAccess, const MemoryLocation &) {
2112 if (auto *Use = dyn_cast<MemoryUseOrDef>(StartingAccess))
2113 return Use->getDefiningAccess();
2114 return StartingAccess;
2117 void MemoryPhi::deleteMe(DerivedUser *Self) {
2118 delete static_cast<MemoryPhi *>(Self);
2121 void MemoryDef::deleteMe(DerivedUser *Self) {
2122 delete static_cast<MemoryDef *>(Self);
2125 void MemoryUse::deleteMe(DerivedUser *Self) {
2126 delete static_cast<MemoryUse *>(Self);